Google launches Gemini 3.5 Flash, its most powerful coding AI yet

Google used its annual developer conference in Mountain View, California, to make a blunt bet on the next phase of AI: not chat, but action. The company said Gemini 3.5 Flash is its strongest agentic and coding model yet, built to execute complex tasks with minimal supervision, from building new applications and maintaining codebases to preparing financial documents.

The pitch is as much about economics as capability. Google said Gemini 3.5 Flash outperforms Gemini 3.1 Pro on demanding coding and agentic benchmarks including Terminal-Bench 2.1, GDPval-AA and MCP Atlas, while delivering output tokens per second at a rate four times faster than other frontier models. In some cases, Google said, the model is meant to cost less than half as much as comparable frontier systems. That combination of speed and lower cost is aimed squarely at developers and enterprises deciding whether AI can move from experimentation to production.

Google is pushing the model straight into products that already have scale. Gemini 3.5 Flash is rolling out globally as the default model in the Gemini app and in AI Mode in Google Search. It is also being used in Google’s updated Antigravity agent-development environment, where the company says it supports collaborative subagents, and in Gemini Spark, the new personal AI agent built on 3.5 Flash. Google said the broader Gemini 3.5 family is designed around “frontier intelligence with action,” a phrase that captures the company’s effort to position agents as the center of its AI strategy.

The launch also shows how quickly Google is moving beyond the chatbot era. Earlier Gemini 3 releases had already pushed more agentic features into Search and developer tools, but 3.5 Flash goes further by tying benchmark gains to concrete workflows and by embedding the model inside Google’s consumer and developer stack at once. Google DeepMind and Google AI Studio have both been central to the rollout, as the company tries to make agentic AI feel less like a demo and more like infrastructure.

The harder test now is execution. A model that can act on behalf of users has to earn trust on reliability, permissions, security and cost, not just benchmark scores. Google’s new model suggests it believes the business case is moving in its favor: if AI can do the work, not just describe it, the market for software, search and enterprise automation gets much bigger.